Le 24/11/2023 à 08:51, John Hearns a écrit :
Good question. Maybe not an answer referring to hwloc.When managing a large NUMA machine, SGI UV, I ran the OS processes in a boot cpuset which was restricted to (AFAIR) the first 8 Cpus. On Intel architecures with E and P cores could we think of running OS on E cores only and having the batch system schedule compute tasks on P cores?
That's certainly possible. Linux has things like isolcpus to force isolate some cores away from the OS tasks, should work for these platforms too (by the way, it's also for ARM big.LITTLE platforms running Linux, including Apple M1, etc).
However, keep in mind that splitting P+E CPUs is not like splitting NUMA platforms: isolating NUMA node #0 on SGI left tons of cores available for HPC tasks on NUMA nodes. Current P+E from Intel usually have more E than P, and several models are even 2P+8E, that would be a lot of E-cores for the OS and very few P-cores for real apps. Your idea would apply better if we rather had 2E+8P but that's not the trend.
Things might be more interesting with MeteorLake which (according to https://www.hardwaretimes.com/intel-14th-gen-meteor-lake-cpu-cores-almost-identical-to-13th-gen-its-a-tic/) has P+E as usual but also 2 "low-power E" on the side. There, you could put the OS on those 2 Low-Power E.
By the way, the Linux scheduler is supposed to get enhanced to automatically find out which tasks to put on P and E core but they've been discussing things for a long time and it's hard to know what's actually working well already.
Brice
OpenPGP_signature
Description: OpenPGP digital signature
_______________________________________________ hwloc-users mailing list hwloc-users@lists.open-mpi.org https://lists.open-mpi.org/mailman/listinfo/hwloc-users
